As media creation evolves, understanding where digital content comes from and how it was created is increasingly important. In this discussion, Pushmeet Kohli, VP of Science and Strategic Initiatives, Google DeepMind, and Hany Farid, Professor at Dartmouth College and founder of GetReal Security, explore how technology can provide clearer context around online media, the role of open standards, the importance of cross-industry collaboration, and the development of SynthID—Google DeepMind’s imperceptible watermarking and detection technology.
Generative AI makes creating realistic content easy, but verifying its authenticity hard. How is this changing how people evaluate what they see, and what responsibility do tech companies share?
Pushmeet: The capabilities of our generative AI models are enabling so many creative applications, but people need to be able to trust what they see online. When photorealistic AI-generated content proliferates, with no way of checking how it’s been created – that is to say, its provenance – it erodes trust.
We believe that innovation and responsibility should never be a trade-off. As a frontier AI lab, we have a duty to protect the information ecosystem with transparency tools like SynthID. Just as seatbelts come standard in cars, digital verification should be available to everyone by default.
Hany: For centuries, we could trust a photo because faking one took immense time, skill, and money. Generative AI removes that friction. Another risk isn't just that people believe fakes, but that inconvenient truths get dismissed as fakes.
People shouldn't have to guess or inspect pixels; platforms and tool creators must build verifiable, open signals directly into the products where people encounter media. Think of it like grocery shopping: we don't ask consumers to run lab tests on their food; we require manufacturers to put clear labels on the package. The companies building and distributing AI tools need to build that same transparency into the technology we use every day – using open standards that work everywhere, not just in their own apps.
We believe that innovation and responsibility should never be a trade-off. As a frontier AI lab, we have a duty to protect the information ecosystem with transparency tools like SynthID. Just as seatbelts come standard in cars, digital verification should be available to everyone by default.
Why aren't traditional detection tools enough on their own anymore?
Hany: Provenance answers two simple questions: where did a file come from, and what changed along the way? It provides a clear chain of custody – showing if an image came from a specific camera or AI model, and whether it was cropped or edited.
Critically, provenance tracks origin, not truthfulness; an authentic photo can still be paired with a false caption. While traditional post-hoc forensic detection remains valuable, it is inherently reactive. Provenance builds accountability right at the point of creation and distribution.
Pushmeet: Many GenAI models are specially trained to ensure that their outputs are as aligned with real data as possible. This leads naturally to the outcome that as models advance, it becomes harder and harder to distinguish their outputs from real images or videos. This is why we decided that the GenAI model outputs should be modified in a way that doesn’t affect their quality but allows detection by users.
SynthID adds an invisible marker to AI-generated content without altering quality. In simple terms, how does that work?
Pushmeet: SynthID weaves a digital watermark directly into AI-generated content as it’s being created. There are various ways to add information about provenance to a piece of media, such as by adding external metadata, but these can often be easily stripped away. Instead, SynthID makes tiny, imperceptible tweaks to the pixels in an image, the sounds in an audio clip, or the word choices in text. You won't notice any change in quality, but our detection tools can spot the pattern instantly.
Launched in 2023, SynthID has grown from a research experiment into a platform that watermarks text, audio, image, and video. It also makes verification simple: you can check if content was made or edited with AI directly in the Gemini app, Search, and Chrome. To date, SynthID has watermarked over 180 billion images and videos and 240,000 years of audio.
How do watermarks like SynthID, open industry standards like C2PA Content Credentials (a label with a file's history), and other forensic tools work together to give people a clearer picture of what they're looking at?
Hany: Each approach solves a different part of the puzzle. SynthID watermarks live inside the media itself and are designed to survive screenshots, cropping, and compression. C2PA Content Credentials provide a transparent history of edits and tools, though metadata can sometimes be stripped. Forensic tools require no opt-in and help catch bad actors who evade watermarks, but they require more time and computing power.
Together, these layers back each other up. If a screenshot strips metadata, the watermark survives; if a bad actor bypasses a watermark, forensic analysis catches it.
Provenance tracks origin, not truthfulness; an authentic photo can still be paired with a false caption. While traditional post-hoc forensic detection remains valuable, it is inherently reactive. Provenance builds accountability right at the point of creation and distribution.
Techniques to bypass watermarks continue to evolve. How is Google addressing these threats, and why is industry collaboration necessary?
Pushmeet: There is no single defense that is permanent or perfect in cybersecurity. We built SynthID to withstand heavy compression, cropping, resizing, and filters. But technical durability is only half the battle. Because media travels across different platforms, bad actors exploit gaps where verification tools aren’t supported.
To close these gaps, we open-sourced SynthID's text watermarking, launched the AI Content Detection API on Google Cloud to detect generative images, and collaborated with companies like Apple, OpenAI, NVIDIA, and Kakao to use SynthID to detect generative media. We are also constantly updating SynthID to respond to a wider array of threats, such as biological molecules. Effective watermarking also makes digital forensics more feasible, because tampering with a watermark leaves detectable traces. This collective, holistic approach to defense raises security for everyone.
Looking ahead 3-5 years, what needs to happen so people can easily know where a piece of media came from? What is your vision for a truly transparent digital ecosystem?
Pushmeet: Provenance verification must become effortless and invisible, just like the security protocols that protect modern web browsers.
There should be clear and trustworthy signals directly within your feeds, computer operating systems, and communication apps that provide you with context about the content you encounter. We’re already integrating C2PA Content Credentials into devices like Pixel phones, but real progress requires close collaboration across industry, policymakers, and civil society.
Hany: We need four things:
- Hardware integration: Provenance must start at the camera lens to verify authentic media, not just mark AI content.
- Platform preservation: Social platforms and apps must stop stripping provenance data upon upload.
- Open interoperability: Verification must work seamlessly across all apps and devices.
- Simple interfaces: People need clear, instant indicators they can understand in seconds.
Everyday people shouldn't have to act like forensic investigators just to browse the internet. That burden belongs with the technology itself. By building simple, built-in verification tools into our platforms, we can help people become more informed about the content they encounter online.